Performance-optimized hierarchical models predict neural responses in higher visual cortex

Representational Distance Learning for Deep Neural Networks

Deep neural networks (DNNs) provide useful models of visual representational transformations. We present a method that enables a DNN (student) to learn from the internal representational spaces of a reference model (teacher), which could be another DNN or, in the future, a biological brain. Representational spaces of the student and the teacher are characterized by representational distance matrices (RDMs). We propose representational distance learning (RDL), a stochastic gradient descent method that drives the RDMs of the student to approximate the RDMs of the teacher. We demonstrate that RDL is competitive with other transfer learning techniques for two publicly available benchmark computer vision datasets (MNIST and CIFAR-100), while allowing for architectural differences between student and teacher. By pulling the student's RDMs toward those of the teacher, RDL significantly improved visual classification performance when compared to baseline networks that did not use transfer learning. In the future, RDL may enable combined supervised training of deep neural networks using task constraints (e.g., images and category labels) and constraints from brain-activity measurements, so as to build models that replicate the internal representational spaces of biological brains.

Building machines that learn and think like people

Recent progress in artificial intelligence has renewed interest in building systems that learn and think like people. Many advances have come from using deep neural networks trained end-to-end in tasks such as object recognition, video games, and board games, achieving performance that equals or even beats that of humans in some respects. Despite their biological inspiration and performance achievements, these systems differ from human intelligence in crucial ways. We review progress in cognitive science suggesting that truly human-like learning and thinking machines will have to reach beyond current engineering trends in both what they learn and how they learn it. Specifically, we argue that these machines should (1) build causal models of the world that support explanation and understanding, rather than merely solving pattern recognition problems; (2) ground learning in intuitive theories of physics and psychology to support and enrich the knowledge that is learned; and (3) harness compositionality and learning-to-learn to rapidly acquire and generalize knowledge to new tasks and situations. We suggest concrete challenges and promising routes toward these goals that can combine the strengths of recent neural network advances with more structured cognitive models.

Modeling the N400 ERP component as transient semantic over-activation within a neural network model of word comprehension

The study of the N400 event-related brain potential has provided fundamental insights into the nature of real-time comprehension processes, and its amplitude is modulated by a wide variety of stimulus and context factors. It is generally thought to reflect the difficulty of semantic access, but formulating a precise characterization of this process has proved difficult. Laszlo and colleagues (Laszlo & Plaut, 2012; Laszlo & Armstrong, 2014) used physiologically constrained neural networks to model the N400 as transient over-activation within semantic representations, arising as a consequence of the distribution of excitation and inhibition within and between cortical areas. The current work extends this approach to successfully model effects on both N400 amplitudes and behavior of word frequency, semantic richness, repetition, semantic and associative priming, and orthographic neighborhood size. The account is argued to be preferable to one based on "implicit semantic prediction error" (Rabovsky & McRae, 2014) for a number of reasons, the most fundamental of which is that the current model actually produces N400-like waveforms in its real-time activation dynamics.

Random synaptic feedback weights support error backpropagation for deep learning

The brain processes information through multiple layers of neurons. This deep architecture is representationally powerful, but complicates learning because it is difficult to identify the responsible neurons when a mistake is made. In machine learning, the backpropagation algorithm assigns blame by multiplying error signals with all the synaptic weights on each neuron's axon and further downstream. However, this involves a precise, symmetric backward connectivity pattern, which is thought to be impossible in the brain. Here we demonstrate that this strong architectural constraint is not required for effective error propagation. We present a surprisingly simple mechanism that assigns blame by multiplying errors by even random synaptic weights. This mechanism can transmit teaching signals across multiple layers of neurons and performs as effectively as backpropagation on a variety of tasks. Our results help reopen questions about how the brain could use error signals and dispel long-held assumptions about algorithmic constraints on learning.

Neurophysiological Organization of the Middle Face Patch in Macaque Inferior Temporal Cortex

While early cortical visual areas contain fine scale spatial organization of neuronal properties, such as orientation preference, the spatial organization of higher-level visual areas is less well understood. The fMRI demonstration of face-preferring regions in human ventral cortex and monkey inferior temporal cortex ("face patches") raises the question of how neural selectivity for faces is organized. Here, we targeted hundreds of spatially registered neural recordings to the largest fMRI-identified face-preferring region in monkeys, the middle face patch (MFP), and show that the MFP contains a graded enrichment of face-preferring neurons. At its center, as much as 93% of the sites we sampled responded twice as strongly to faces than to nonface objects. We estimate the maximum neurophysiological size of the MFP to be ∼6 mm in diameter, consistent with its previously reported size under fMRI. Importantly, face selectivity in the MFP varied strongly even between neighboring sites. Additionally, extremely face-selective sites were ∼40 times more likely to be present inside the MFP than outside. These results provide the first direct quantification of the size and neural composition of the MFP by showing that the cortical tissue localized to the fMRI defined region consists of a very high fraction of face-preferring sites near its center, and a monotonic decrease in that fraction along any radial spatial axis.

SIGNIFICANCE STATEMENT

The underlying organization of neurons that give rise to the large spatial regions of activity observed with fMRI is not well understood. Neurophysiological studies that have targeted the fMRI identified face patches in monkeys have provided evidence for both large-scale clustering and a heterogeneous spatial organization. Here we used a novel x-ray imaging system to spatially map the responses of hundreds of sites in and around the middle face patch. We observed that face-selective signal localized to the middle face patch was characterized by a gradual spatial enrichment. Furthermore, strongly face-selective sites were ∼40 times more likely to be found inside the patch than outside of the patch.

Neural Quadratic Discriminant Analysis: Nonlinear Decoding with V1-Like Computation

Linear-nonlinear (LN) models and their extensions have proven successful in describing transformations from stimuli to spiking responses of neurons in early stages of sensory hierarchies. Neural responses at later stages are highly nonlinear and have generally been better characterized in terms of their decoding performance on prespecified tasks. Here we develop a biologically plausible decoding model for classification tasks, that we refer to as neural quadratic discriminant analysis (nQDA). Specifically, we reformulate an optimal quadratic classifier as an LN-LN computation, analogous to "subunit" encoding models that have been used to describe responses in retina and primary visual cortex. We propose a physiological mechanism by which the parameters of the nQDA classifier could be optimized, using a supervised variant of a Hebbian learning rule. As an example of its applicability, we show that nQDA provides a better account than many comparable alternatives for the transformation between neural representations in two high-level brain areas recorded as monkeys performed a visual delayed-match-to-sample task.

Making Sense of Real-World Scenes

To interact with the world, we have to make sense of the continuous sensory input conveying information about our environment. A recent surge of studies has investigated the processes enabling scene understanding, using increasingly complex stimuli and sophisticated analyses to highlight the visual features and brain regions involved. However, there are two major challenges to producing a comprehensive framework for scene understanding. First, scene perception is highly dynamic, subserving multiple behavioral goals. Second, a multitude of different visual properties co-occur across scenes and may be correlated or independent. We synthesize the recent literature and argue that for a complete view of scene understanding, it is necessary to account for both differing observer goals and the contribution of diverse scene properties.

Nonlinear Hebbian Learning as a Unifying Principle in Receptive Field Formation

The development of sensory receptive fields has been modeled in the past by a variety of models including normative models such as sparse coding or independent component analysis and bottom-up models such as spike-timing dependent plasticity or the Bienenstock-Cooper-Munro model of synaptic plasticity. Here we show that the above variety of approaches can all be unified into a single common principle, namely nonlinear Hebbian learning. When nonlinear Hebbian learning is applied to natural images, receptive field shapes were strongly constrained by the input statistics and preprocessing, but exhibited only modest variation across different choices of nonlinearities in neuron models or synaptic plasticity rules. Neither overcompleteness nor sparse network activity are necessary for the development of localized receptive fields. The analysis of alternative sensory modalities such as auditory models or V2 development lead to the same conclusions. In all examples, receptive fields can be predicted a priori by reformulating an abstract model as nonlinear Hebbian learning. Thus nonlinear Hebbian learning and natural statistics can account for many aspects of receptive field formation across models and sensory modalities.

The question of how the brain self-organizes to develop precisely tuned neurons has puzzled neuroscientists at least since the discoveries of Hubel and Wiesel. In the past decades, a variety of theories and models have been proposed to describe receptive field formation, notably V1 simple cells, from natural inputs. We cut through the jungle of candidate explanations by demonstrating that in fact a single principle is sufficient to explain receptive field development. Our results follow from two major insights. First, we show that many representative models of sensory development are in fact implementing variations of a common principle: nonlinear Hebbian learning. Second, we reveal that nonlinear Hebbian learning is sufficient for receptive field formation through sensory inputs. The surprising result is that our findings are robust of specific details of a model, and allows for robust predictions on the learned receptive fields. Nonlinear Hebbian learning is therefore general in two senses: it applies to many models developed by theoreticians, and to many sensory modalities studied by experimental neuroscientists.

Deep Networks Can Resemble Human Feed-forward Vision in Invariant Object Recognition

Deep convolutional neural networks (DCNNs) have attracted much attention recently, and have shown to be able to recognize thousands of object categories in natural image databases. Their architecture is somewhat similar to that of the human visual system: both use restricted receptive fields, and a hierarchy of layers which progressively extract more and more abstracted features. Yet it is unknown whether DCNNs match human performance at the task of view-invariant object recognition, whether they make similar errors and use similar representations for this task, and whether the answers depend on the magnitude of the viewpoint variations. To investigate these issues, we benchmarked eight state-of-the-art DCNNs, the HMAX model, and a baseline shallow model and compared their results to those of humans with backward masking. Unlike in all previous DCNN studies, we carefully controlled the magnitude of the viewpoint variations to demonstrate that shallow nets can outperform deep nets and humans when variations are weak. When facing larger variations, however, more layers were needed to match human performance and error distributions, and to have representations that are consistent with human behavior. A very deep net with 18 layers even outperformed humans at the highest variation level, using the most human-like representations.

How Invariant Feature Selectivity Is Achieved in Cortex

Parsing the visual scene into objects is paramount to survival. Yet, how this is accomplished by the nervous system remains largely unknown, even in the comparatively well understood visual system. It is especially unclear how detailed peripheral signal representations are transformed into the object-oriented representations that are independent of object position and are provided by the final stages of visual processing. This perspective discusses advances in computational algorithms for fitting large-scale models that make it possible to reconstruct the intermediate steps of visual processing based on neural responses to natural stimuli. In particular, it is now possible to characterize how different types of position invariance, such as local (also known as phase invariance) and more global, are interleaved with nonlinear operations to allow for coding of curved contours. Neurons in the mid-level visual area V4 exhibit selectivity to pairs of even- and odd-symmetric profiles along curved contours. Such pairing is reminiscent of the response properties of complex cells in the primary visual cortex (V1) and suggests specific ways in which V1 signals are transformed within subsequent visual cortical areas. These examples illustrate that large-scale models fitted to neural responses to natural stimuli can provide generative models of successive stages of sensory processing.

A null model for cortical representations with grandmothers galore

There has been extensive discussion in the literature about the extent to which cortical representations can be described as localist or distributed. Here we discuss a simple null model that encompasses a family of related architectures describing the transformation of signals throughout the parts of the visual system involved in object recognition. This family of models constitutes a rigorous first approximation to explain the neurophysiological properties of ventral visual cortex. This null model contains both distributed and local representations throughout the entire hierarchy of computations and the responses of individual units are meaningful and interpretable when encoding is adequately defined for each computational stage.

Early Visual Cortex as a Multiscale Cognitive Blackboard

Neurons in early visual cortical areas not only represent incoming visual information but are also engaged by higher level cognitive processes, including attention, working memory, imagery, and decision-making. Are these cognitive effects an epiphenomenon or are they functionally relevant for these mental operations? We review evidence supporting the hypothesis that the modulation of activity in early visual areas has a causal role in cognition. The modulatory influences allow the early visual cortex to act as a multiscale cognitive blackboard for read and write operations by higher visual areas, which can thereby efficiently exchange information. This blackboard architecture explains how the activity of neurons in the early visual cortex contributes to scene segmentation and working memory, and relates to the subject's inferences about the visual world. The architecture also has distinct advantages for the processing of visual routines that rely on a number of sequentially executed processing steps.

Serial grouping of 2D-image regions with object-based attention in humans

After an initial stage of local analysis within the retina and early visual pathways, the human visual system creates a structured representation of the visual scene by co-selecting image elements that are part of behaviorally relevant objects. The mechanisms underlying this perceptual organization process are only partially understood. We here investigate the time-course of perceptual grouping of two-dimensional image-regions by measuring the reaction times of human participants and report that it is associated with the gradual spread of object-based attention. Attention spreads fastest over large and homogeneous areas and is slowed down at locations that require small-scale processing. We find that the time-course of the object-based selection process is well explained by a 'growth-cone' model, which selects surface elements in an incremental, scale-dependent manner. We discuss how the visual cortical hierarchy can implement this scale-dependent spread of object-based attention, leveraging the different receptive field sizes in distinct cortical areas.

DOI:http://dx.doi.org/10.7554/eLife.14320.001

When we look at an object, we perceive it as a whole. However, this is not how the brain processes objects. Instead, cells at early stages of the visual system respond selectively to single features of the object, such as edges. Moreover, each cell responds to its target feature in only a small region of space known as its receptive field. At higher levels of the visual system, cells respond to more complex features: angles rather than edges, for example. The receptive fields of the cells are also larger. For us to see an object, the brain must therefore 'stitch' together diverse features into a unified impression.

This process is termed perceptual grouping. But how does it work? Jeurissen et al. hypothesized that this process depends on the visual system’s attention spreading over a region in the image occupied by an object, and that the speed of the process will depend on the size of the receptive fields involved. If an image region is narrow, the visual system must recruit cells with small receptive fields to process the individual features. Grouping will therefore be slow. By contrast, if the object consists of large uniform areas lacking in detail, grouping should be fast. These assumptions give rise to a model called the “growth-conemodel”, which makes a number of specific predictions about reaction times during perceptual grouping.

Jeurissen et al. tested the growth-cone model’s predictions by measuring the speed of perceptual grouping in 160 human volunteers. These volunteers looked at an image made up of two simple shapes, and reported whether two dots fell on the same or different shapes. The results supported the growth-cone model. People were able to group large and uniform areas quickly, but were slower for narrow areas. Grouping also took more time when the distance between the dots increased. Hence, perceptual grouping of everyday objects calls on a step-by-step process that resembles solving a small maze.

The results also revealed that perceptual grouping of simple shapes relies on the spreading of visual attention over the relevant object. Furthermore, the data support the hypothesis that perceptual grouping makes use of the different sizes of receptive fields at various levels of the visual system. Further research will be needed to translate these findings to the more complex natural scenes we encounter in our daily lives.

DOI:http://dx.doi.org/10.7554/eLife.14320.002

Comparison of deep neural networks to spatio-temporal cortical dynamics of human visual object recognition reveals hierarchical correspondence

The complex multi-stage architecture of cortical visual pathways provides the neural basis for efficient visual object recognition in humans. However, the stage-wise computations therein remain poorly understood. Here, we compared temporal (magnetoencephalography) and spatial (functional MRI) visual brain representations with representations in an artificial deep neural network (DNN) tuned to the statistics of real-world visual recognition. We showed that the DNN captured the stages of human visual processing in both time and space from early visual areas towards the dorsal and ventral streams. Further investigation of crucial DNN parameters revealed that while model architecture was important, training on real-world categorization was necessary to enforce spatio-temporal hierarchical relationships with the brain. Together our results provide an algorithmically informed view on the spatio-temporal dynamics of visual object recognition in the human visual brain.

Category-Selectivity in Human Visual Cortex Follows Cortical Topology: A Grouped icEEG Study

Neuroimaging studies suggest that category-selective regions in higher-order visual cortex are topologically organized around specific anatomical landmarks: the mid-fusiform sulcus (MFS) in the ventral temporal cortex (VTC) and lateral occipital sulcus (LOS) in the lateral occipital cortex (LOC). To derive precise structure-function maps from direct neural signals, we collected intracranial EEG (icEEG) recordings in a large human cohort (n = 26) undergoing implantation of subdural electrodes. A surface-based approach to grouped icEEG analysis was used to overcome challenges from sparse electrode coverage within subjects and variable cortical anatomy across subjects. The topology of category-selectivity in bilateral VTC and LOC was assessed for five classes of visual stimuli-faces, animate non-face (animals/body-parts), places, tools, and words-using correlational and linear mixed effects analyses. In the LOC, selectivity for living (faces and animate non-face) and non-living (places and tools) classes was arranged in a ventral-to-dorsal axis along the LOS. In the VTC, selectivity for living and non-living stimuli was arranged in a latero-medial axis along the MFS. Written word-selectivity was reliably localized to the intersection of the left MFS and the occipito-temporal sulcus. These findings provide direct electrophysiological evidence for topological information structuring of functional representations within higher-order visual cortex.

Selectivity and tolerance for visual texture in macaque V2

As information propagates along the ventral visual hierarchy, neuronal responses become both more specific for particular image features and more tolerant of image transformations that preserve those features. Here, we present evidence that neurons in area V2 are selective for local statistics that occur in natural visual textures, and tolerant of manipulations that preserve these statistics. Texture stimuli were generated by sampling from a statistical model, with parameters chosen to match the parameters of a set of visually distinct natural texture images. Stimuli generated with the same statistics are perceptually similar to each other despite differences, arising from the sampling process, in the precise spatial location of features. We assessed the accuracy with which these textures could be classified based on the responses of V1 and V2 neurons recorded individually in anesthetized macaque monkeys. We also assessed the accuracy with which particular samples could be identified, relative to other statistically matched samples. For populations of up to 100 cells, V1 neurons supported better performance in the sample identification task, whereas V2 neurons exhibited better performance in texture classification. Relative to V1, the responses of V2 show greater selectivity and tolerance for the representation of texture statistics.

Temporal dynamics of visual category representation in the macaque inferior temporal cortex

Object categories are recognized at multiple levels of hierarchical abstractions. Psychophysical studies have shown a more rapid perceptual access to the mid-level category information (e.g., human faces) than the higher (superordinate; e.g., animal) or the lower (subordinate; e.g., face identity) level. Mid-level category members share many features, whereas few features are shared among members of different mid-level categories. To understand better the neural basis of expedited access to mid-level category information, we examined neural responses of the inferior temporal (IT) cortex of macaque monkeys viewing a large number of object images. We found an earlier representation of mid-level categories in the IT population and single-unit responses compared with superordinate- and subordinate-level categories. The short-latency representation of mid-level category information shows that visual cortex first divides the category shape space at its sharpest boundaries, defined by high/low within/between-group similarity. This short-latency, mid-level category boundary map may be a prerequisite for representation of other categories at more global and finer scales.

Learning the 3-D structure of objects from 2-D views depends on shape, not format

Humans can learn to recognize new objects just from observing example views. However, it is unknown what structural information enables this learning. To address this question, we manipulated the amount of structural information given to subjects during unsupervised learning by varying the format of the trained views. We then tested how format affected participants' ability to discriminate similar objects across views that were rotated 90° apart. We found that, after training, participants' performance increased and generalized to new views in the same format. Surprisingly, the improvement was similar across line drawings, shape from shading, and shape from shading + stereo even though the latter two formats provide richer depth information compared to line drawings. In contrast, participants' improvement was significantly lower when training used silhouettes, suggesting that silhouettes do not have enough information to generate a robust 3-D structure. To test whether the learned object representations were format-specific or format-invariant, we examined if learning novel objects from example views transfers across formats. We found that learning objects from example line drawings transferred to shape from shading and vice versa. These results have important implications for theories of object recognition because they suggest that (a) learning the 3-D structure of objects does not require rich structural cues during training as long as shape information of internal and external features is provided and (b) learning generates shape-based object representations independent of the training format.

Deep Neural Networks as a Computational Model for Human Shape Sensitivity

Theories of object recognition agree that shape is of primordial importance, but there is no consensus about how shape might be represented, and so far attempts to implement a model of shape perception that would work with realistic stimuli have largely failed. Recent studies suggest that state-of-the-art convolutional ‘deep’ neural networks (DNNs) capture important aspects of human object perception. We hypothesized that these successes might be partially related to a human-like representation of object shape. Here we demonstrate that sensitivity for shape features, characteristic to human and primate vision, emerges in DNNs when trained for generic object recognition from natural photographs. We show that these models explain human shape judgments for several benchmark behavioral and neural stimulus sets on which earlier models mostly failed. In particular, although never explicitly trained for such stimuli, DNNs develop acute sensitivity to minute variations in shape and to non-accidental properties that have long been implicated to form the basis for object recognition. Even more strikingly, when tested with a challenging stimulus set in which shape and category membership are dissociated, the most complex model architectures capture human shape sensitivity as well as some aspects of the category structure that emerges from human judgments. As a whole, these results indicate that convolutional neural networks not only learn physically correct representations of object categories but also develop perceptually accurate representational spaces of shapes. An even more complete model of human object representations might be in sight by training deep architectures for multiple tasks, which is so characteristic in human development.

Shape plays an important role in object recognition. Despite years of research, no models of vision could account for shape understanding as found in human vision of natural images. Given recent successes of deep neural networks (DNNs) in object recognition, we hypothesized that DNNs might in fact learn to capture perceptually salient shape dimensions. Using a variety of stimulus sets, we demonstrate here that the output layers of several DNNs develop representations that relate closely to human perceptual shape judgments. Surprisingly, such sensitivity to shape develops in these models even though they were never explicitly trained for shape processing. Moreover, we show that these models also represent categorical object similarity that follows human semantic judgments, albeit to a lesser extent. Taken together, our results bring forward the exciting idea that DNNs capture not only objective dimensions of stimuli, such as their category, but also their subjective, or perceptual, aspects, such as shape and semantic similarity as judged by humans.

The development of newborn object recognition in fast and slow visual worlds

Object recognition is central to perception and cognition. Yet relatively little is known about the environmental factors that cause invariant object recognition to emerge in the newborn brain. Is this ability a hardwired property of vision? Or does the development of invariant object recognition require experience with a particular kind of visual environment? Here, we used a high-throughput controlled-rearing method to examine whether newborn chicks (Gallus gallus) require visual experience with slowly changing objects to develop invariant object recognition abilities. When newborn chicks were raised with a slowly rotating virtual object, the chicks built invariant object representations that generalized across novel viewpoints and rotation speeds. In contrast, when newborn chicks were raised with a virtual object that rotated more quickly, the chicks built viewpoint-specific object representations that failed to generalize to novel viewpoints and rotation speeds. Moreover, there was a direct relationship between the speed of the object and the amount of invariance in the chick's object representation. Thus, visual experience with slowly changing objects plays a critical role in the development of invariant object recognition. These results indicate that invariant object recognition is not a hardwired property of vision, but is learned rapidly when newborns encounter a slowly changing visual world.

A specialized face-processing model inspired by the organization of monkey face patches explains several face-specific phenomena observed in humans

Converging reports indicate that face images are processed through specialized neural networks in the brain –i.e. face patches in monkeys and the fusiform face area (FFA) in humans. These studies were designed to find out how faces are processed in visual system compared to other objects. Yet, the underlying mechanism of face processing is not completely revealed. Here, we show that a hierarchical computational model, inspired by electrophysiological evidence on face processing in primates, is able to generate representational properties similar to those observed in monkey face patches (posterior, middle and anterior patches). Since the most important goal of sensory neuroscience is linking the neural responses with behavioral outputs, we test whether the proposed model, which is designed to account for neural responses in monkey face patches, is also able to predict well-documented behavioral face phenomena observed in humans. We show that the proposed model satisfies several cognitive face effects such as: composite face effect and the idea of canonical face views. Our model provides insights about the underlying computations that transfer visual information from posterior to anterior face patches.

Enhanced learning of natural visual sequences in newborn chicks

To what extent are newborn brains designed to operate over natural visual input? To address this question, we used a high-throughput controlled-rearing method to examine whether newborn chicks (Gallus gallus) show enhanced learning of natural visual sequences at the onset of vision. We took the same set of images and grouped them into either natural sequences (i.e., sequences showing different viewpoints of the same real-world object) or unnatural sequences (i.e., sequences showing different images of different real-world objects). When raised in virtual worlds containing natural sequences, newborn chicks developed the ability to recognize familiar images of objects. Conversely, when raised in virtual worlds containing unnatural sequences, newborn chicks' object recognition abilities were severely impaired. In fact, the majority of the chicks raised with the unnatural sequences failed to recognize familiar images of objects despite acquiring over 100 h of visual experience with those images. Thus, newborn chicks show enhanced learning of natural visual sequences at the onset of vision. These results indicate that newborn brains are designed to operate over natural visual input.

Typicality sharpens category representations in object-selective cortex

The purpose of categorization is to identify generalizable classes of objects whose members can be treated equivalently. Within a category, however, some exemplars are more representative of that concept than others. Despite long-standing behavioral effects, little is known about how typicality influences the neural representation of real-world objects from the same category. Using fMRI, we showed participants 64 subordinate object categories (exemplars) grouped into 8 basic categories. Typicality for each exemplar was assessed behaviorally and we used several multi-voxel pattern analyses to characterize how typicality affects the pattern of responses elicited in early visual and object-selective areas: V1, V2, V3v, hV4, LOC. We found that in LOC, but not in early areas, typical exemplars elicited activity more similar to the central category tendency and created sharper category boundaries than less typical exemplars, suggesting that typicality enhances within-category similarity and between-category dissimilarity. Additionally, we uncovered a brain region (cIPL) where category boundaries favor less typical categories. Our results suggest that typicality may constitute a previously unexplored principle of organization for intra-category neural structure and, furthermore, that this representation is not directly reflected in image features describing natural input, but rather built by the visual system at an intermediate processing stage.

Stimulus features coded by single neurons of a macaque body category selective patch

Body category-selective regions of the primate temporal cortex respond to images of bodies, but it is unclear which fragments of such images drive single neurons' responses in these regions. Here we applied the Bubbles technique to the responses of single macaque middle superior temporal sulcus (midSTS) body patch neurons to reveal the image fragments the neurons respond to. We found that local image fragments such as extremities (limbs), curved boundaries, and parts of the torso drove the large majority of neurons. Bubbles revealed the whole body in only a few neurons. Neurons coded the features in a manner that was tolerant to translation and scale changes. Most image fragments were excitatory but for a few neurons both inhibitory and excitatory fragments (opponent coding) were present in the same image. The fragments we reveal here in the body patch with Bubbles differ from those suggested in previous studies of face-selective neurons in face patches. Together, our data indicate that the majority of body patch neurons respond to local image fragments that occur frequently, but not exclusively, in bodies, with a coding that is tolerant to translation and scale. Overall, the data suggest that the body category selectivity of the midSTS body patch depends more on the feature statistics of bodies (e.g., extensions occur more frequently in bodies) than on semantics (bodies as an abstract category).

Explicit information for category-orthogonal object properties increases along the ventral stream

Extensive research has revealed that the ventral visual stream hierarchically builds a robust representation for supporting visual object categorization tasks. We systematically explored the ability of multiple ventral visual areas to support a variety of 'category-orthogonal' object properties such as position, size and pose. For complex naturalistic stimuli, we found that the inferior temporal (IT) population encodes all measured category-orthogonal object properties, including those properties often considered to be low-level features (for example, position), more explicitly than earlier ventral stream areas. We also found that the IT population better predicts human performance patterns across properties. A hierarchical neural network model based on simple computational principles generates these same cross-area patterns of information. Taken together, our empirical results support the hypothesis that all behaviorally relevant object properties are extracted in concert up the ventral visual hierarchy, and our computational model explains how that hierarchy might be built.

Visual information representation and rapid-scene categorization are simultaneous across cortex: An MEG study

Perceiving the visual world around us requires the brain to represent the features of stimuli and to categorize the stimulus based on these features. Incorrect categorization can result either from errors in visual representation or from errors in processes that lead to categorical choice. To understand the temporal relationship between the neural signatures of such systematic errors, we recorded whole-scalp magnetoencephalography (MEG) data from human subjects performing a rapid-scene categorization task. We built scene category decoders based on (1) spatiotemporally resolved neural activity, (2) spatial envelope (SpEn) image features, and (3) behavioral responses. Using confusion matrices, we tracked how well the pattern of errors from neural decoders could be explained by SpEn decoders and behavioral errors, over time and across cortical areas. Across the visual cortex and the medial temporal lobe, we found that both SpEn and behavioral errors explained unique variance in the errors of neural decoders. Critically, these effects were nearly simultaneous, and most prominent between 100 and 250ms after stimulus onset. Thus, during rapid-scene categorization, neural processes that ultimately result in behavioral categorization are simultaneous and co-localized with neural processes underlying visual information representation.

Atoms of recognition in human and computer vision

Discovering the visual features and representations used by the brain to recognize objects is a central problem in the study of vision. Recently, neural network models of visual object recognition, including biological and deep network models, have shown remarkable progress and have begun to rival human performance in some challenging tasks. These models are trained on image examples and learn to extract features and representations and to use them for categorization. It remains unclear, however, whether the representations and learning processes discovered by current models are similar to those used by the human visual system. Here we show, by introducing and using minimal recognizable images, that the human visual system uses features and processes that are not used by current models and that are critical for recognition. We found by psychophysical studies that at the level of minimal recognizable images a minute change in the image can have a drastic effect on recognition, thus identifying features that are critical for the task. Simulations then showed that current models cannot explain this sensitivity to precise feature configurations and, more generally, do not learn to recognize minimal images at a human level. The role of the features shown here is revealed uniquely at the minimal level, where the contribution of each feature is essential. A full understanding of the learning and use of such features will extend our understanding of visual recognition and its cortical mechanisms and will enhance the capacity of computational models to learn from visual experience and to deal with recognition and detailed image interpretation.

Using goal-driven deep learning models to understand sensory cortex

Fueled by innovation in the computer vision and artificial intelligence communities, recent developments in computational neuroscience have used goal-driven hierarchical convolutional neural networks (HCNNs) to make strides in modeling neural single-unit and population responses in higher visual cortical areas. In this Perspective, we review the recent progress in a broader modeling context and describe some of the key technical innovations that have supported it. We then outline how the goal-driven HCNN approach can be used to delve even more deeply into understanding the development and organization of sensory cortical processing.

Training Excitatory-Inhibitory Recurrent Neural Networks for Cognitive Tasks: A Simple and Flexible Framework

The ability to simultaneously record from large numbers of neurons in behaving animals has ushered in a new era for the study of the neural circuit mechanisms underlying cognitive functions. One promising approach to uncovering the dynamical and computational principles governing population responses is to analyze model recurrent neural networks (RNNs) that have been optimized to perform the same tasks as behaving animals. Because the optimization of network parameters specifies the desired output but not the manner in which to achieve this output, “trained” networks serve as a source of mechanistic hypotheses and a testing ground for data analyses that link neural computation to behavior. Complete access to the activity and connectivity of the circuit, and the ability to manipulate them arbitrarily, make trained networks a convenient proxy for biological circuits and a valuable platform for theoretical investigation. However, existing RNNs lack basic biological features such as the distinction between excitatory and inhibitory units (Dale’s principle), which are essential if RNNs are to provide insights into the operation of biological circuits. Moreover, trained networks can achieve the same behavioral performance but differ substantially in their structure and dynamics, highlighting the need for a simple and flexible framework for the exploratory training of RNNs. Here, we describe a framework for gradient descent-based training of excitatory-inhibitory RNNs that can incorporate a variety of biological knowledge. We provide an implementation based on the machine learning library Theano, whose automatic differentiation capabilities facilitate modifications and extensions. We validate this framework by applying it to well-known experimental paradigms such as perceptual decision-making, context-dependent integration, multisensory integration, parametric working memory, and motor sequence generation. Our results demonstrate the wide range of neural activity patterns and behavior that can be modeled, and suggest a unified setting in which diverse cognitive computations and mechanisms can be studied.

Cognitive functions arise from the coordinated activity of many interconnected neurons. As neuroscientists increasingly use large datasets of simultaneously recorded neurons to study the brain, one approach that has emerged as a promising tool for interpreting population responses is to analyze model recurrent neural networks (RNNs) that have been optimized to perform the same tasks as recorded animals. Complete access to the activity and connectivity of the circuit, and the ability to manipulate them in arbitrary ways, make trained networks a convenient proxy for biological circuits and a valuable platform for theoretical investigation. However, existing RNNs lack basic biological features that are essential if RNNs are to provide insights into the circuit-level operation of the brain. Moreover, trained networks can achieve the same behavioral performance but differ substantially in their structure and dynamics, highlighting the need for a simple and flexible framework for the exploratory training of RNNs. Here we describe and provide an implementation for such a framework, which we apply to several well-known experimental paradigms that illustrate the diversity of detail that can be modeled. Our work provides a foundation for neuroscientists to harness trained RNNs in their own investigations of the neural basis of cognition.

Conjectures regarding the nonlinear geometry of visual neurons

From the earliest stages of sensory processing, neurons show inherent non-linearities: the response to a complex stimulus is not a sum of the responses to a set of constituent basis stimuli. These non-linearities come in a number of forms and have been explained in terms of a number of functional goals. The family of spatial non-linearities have included interactions that occur both within and outside of the classical receptive field. They include, saturation, cross orientation inhibition, contrast normalization, end-stopping and a variety of non-classical effects. In addition, neurons show a number of facilitatory and invariance related effects such as those exhibited by complex cells (integration across position). Here, we describe an approach that attempts to explain many of the non-linearities under a single geometric framework. In line with Zetzsche and colleagues (e.g., Zetzsche et al., 1999) we propose that many of the principal non-linearities can be described by a geometry where the neural response space has a simple curvature. In this paper, we focus on the geometry that produces both increased selectivity (curving outward) and increased tolerance (curving inward). We demonstrate that overcomplete sparse coding with both low-dimensional synthetic data and high-dimensional natural scene data can result in curvature that is responsible for a variety of different known non-classical effects including end-stopping and gain control. We believe that this approach provides a more fundamental explanation of these non-linearities and does not require that one postulate a variety of explanations (e.g., that gain must be controlled or the ends of lines must be detected). In its standard form, sparse coding does not however, produce invariance/tolerance represented by inward curvature. We speculate on some of the requirements needed to produce such curvature.

Canonical computations of cerebral cortex

The idea that there is a fundamental cortical circuit that performs canonical computations remains compelling though far from proven. Here we review evidence for two canonical operations within sensory cortical areas: a feedforward computation of selectivity; and a recurrent computation of gain in which, given sufficiently strong external input, perhaps from multiple sources, intracortical input largely, but not completely, cancels this external input. This operation leads to many characteristic cortical nonlinearities in integrating multiple stimuli. The cortical computation must combine such local processing with hierarchical processing across areas. We point to important changes in moving from sensory cortex to motor and frontal cortex and the possibility of substantial differences between cortex in rodents vs. species with columnar organization of selectivity.

Why neurons mix: high dimensionality for higher cognition

Neurons often respond to diverse combinations of task-relevant variables. This form of mixed selectivity plays an important computational role which is related to the dimensionality of the neural representations: high-dimensional representations with mixed selectivity allow a simple linear readout to generate a huge number of different potential responses. In contrast, neural representations based on highly specialized neurons are low dimensional and they preclude a linear readout from generating several responses that depend on multiple task-relevant variables. Here we review the conceptual and theoretical framework that explains the importance of mixed selectivity and the experimental evidence that recorded neural representations are high-dimensional. We end by discussing the implications for the design of future experiments.

Neural and Computational Mechanisms of Action Processing: Interaction between Visual and Motor Representations

Action recognition has received enormous interest in the field of neuroscience over the last two decades. In spite of this interest, the knowledge in terms of fundamental neural mechanisms that provide constraints for underlying computations remains rather limited. This fact stands in contrast with a wide variety of speculative theories about how action recognition might work. This review focuses on new fundamental electrophysiological results in monkeys, which provide constraints for the detailed underlying computations. In addition, we review models for action recognition and processing that have concrete mathematical implementations, as opposed to conceptual models. We think that only such implemented models can be meaningfully linked quantitatively to physiological data and have a potential to narrow down the many possible computational explanations for action recognition. In addition, only concrete implementations allow judging whether postulated computational concepts have a feasible implementation in terms of realistic neural circuits.

Are Face and Object Recognition Independent? A Neurocomputational Modeling Exploration

Are face and object recognition abilities independent? Although it is commonly believed that they are, Gauthier et al. [Gauthier, I., McGugin, R. W., Richler, J. J., Herzmann, G., Speegle, M., & VanGulick, A. E. Experience moderates overlap between object and face recognition, suggesting a common ability. Journal of Vision, 14, 7, 2014] recently showed that these abilities become more correlated as experience with nonface categories increases. They argued that there is a single underlying visual ability, v, that is expressed in performance with both face and nonface categories as experience grows. Using the Cambridge Face Memory Test and the Vanderbilt Expertise Test, they showed that the shared variance between Cambridge Face Memory Test and Vanderbilt Expertise Test performance increases monotonically as experience increases. Here, we address why a shared resource across different visual domains does not lead to competition and to an inverse correlation in abilities? We explain this conundrum using our neurocomputational model of face and object processing ["The Model", TM, Cottrell, G. W., & Hsiao, J. H. Neurocomputational models of face processing. In A. J. Calder, G. Rhodes, M. Johnson, & J. Haxby (Eds.), The Oxford handbook of face perception. Oxford, UK: Oxford University Press, 2011]. We model the domain general ability v as the available computational resources (number of hidden units) in the mapping from input to label and experience as the frequency of individual exemplars in an object category appearing during network training. Our results show that, as in the behavioral data, the correlation between subordinate level face and object recognition accuracy increases as experience grows. We suggest that different domains do not compete for resources because the relevant features are shared between faces and objects. The essential power of experience is to generate a "spreading transform" for faces (separating them in representational space) that generalizes to objects that must be individuated. Interestingly, when the task of the network is basic level categorization, no increase in the correlation between domains is observed. Hence, our model predicts that it is the type of experience that matters and that the source of the correlation is in the fusiform face area, rather than in cortical areas that subserve basic level categorization. This result is consistent with our previous modeling elucidating why the FFA is recruited for novel domains of expertise [Tong, M. H., Joyce, C. A., & Cottrell, G. W. Why is the fusiform face area recruited for novel categories of expertise? A neurocomputational investigation. Brain Research, 1202, 14-24, 2008].

Simple Learned Weighted Sums of Inferior Temporal Neuronal Firing Rates Accurately Predict Human Core Object Recognition Performance

To go beyond qualitative models of the biological substrate of object recognition, we ask: can a single ventral stream neuronal linking hypothesis quantitatively account for core object recognition performance over a broad range of tasks? We measured human performance in 64 object recognition tests using thousands of challenging images that explore shape similarity and identity preserving object variation. We then used multielectrode arrays to measure neuronal population responses to those same images in visual areas V4 and inferior temporal (IT) cortex of monkeys and simulated V1 population responses. We tested leading candidate linking hypotheses and control hypotheses, each postulating how ventral stream neuronal responses underlie object recognition behavior. Specifically, for each hypothesis, we computed the predicted performance on the 64 tests and compared it with the measured pattern of human performance. All tested hypotheses based on low- and mid-level visually evoked activity (pixels, V1, and V4) were very poor predictors of the human behavioral pattern. However, simple learned weighted sums of distributed average IT firing rates exactly predicted the behavioral pattern. More elaborate linking hypotheses relying on IT trial-by-trial correlational structure, finer IT temporal codes, or ones that strictly respect the known spatial substructures of IT ("face patches") did not improve predictive power. Although these results do not reject those more elaborate hypotheses, they suggest a simple, sufficient quantitative model: each object recognition task is learned from the spatially distributed mean firing rates (100 ms) of ∼60,000 IT neurons and is executed as a simple weighted sum of those firing rates. Significance statement: We sought to go beyond qualitative models of visual object recognition and determine whether a single neuronal linking hypothesis can quantitatively account for core object recognition behavior. To achieve this, we designed a database of images for evaluating object recognition performance. We used multielectrode arrays to characterize hundreds of neurons in the visual ventral stream of nonhuman primates and measured the object recognition performance of >100 human observers. Remarkably, we found that simple learned weighted sums of firing rates of neurons in monkey inferior temporal (IT) cortex accurately predicted human performance. Although previous work led us to expect that IT would outperform V4, we were surprised by the quantitative precision with which simple IT-based linking hypotheses accounted for human behavior.

Semantic Wavelet-Induced Frequency-Tagging (SWIFT) Periodically Activates Category Selective Areas While Steadily Activating Early Visual Areas

Primate visual systems process natural images in a hierarchical manner: at the early stage, neurons are tuned to local image features, while neurons in high-level areas are tuned to abstract object categories. Standard models of visual processing assume that the transition of tuning from image features to object categories emerges gradually along the visual hierarchy. Direct tests of such models remain difficult due to confounding alteration in low-level image properties when contrasting distinct object categories. When such contrast is performed in a classic functional localizer method, the desired activation in high-level visual areas is typically accompanied with activation in early visual areas. Here we used a novel image-modulation method called SWIFT (semantic wavelet-induced frequency-tagging), a variant of frequency-tagging techniques. Natural images modulated by SWIFT reveal object semantics periodically while keeping low-level properties constant. Using functional magnetic resonance imaging (fMRI), we indeed found that faces and scenes modulated with SWIFT periodically activated the prototypical category-selective areas while they elicited sustained and constant responses in early visual areas. SWIFT and the localizer were selective and specific to a similar extent in activating category-selective areas. Only SWIFT progressively activated the visual pathway from low- to high-level areas, consistent with predictions from standard hierarchical models. We confirmed these results with criterion-free methods, generalizing the validity of our approach and show that it is possible to dissociate neural activation in early and category-selective areas. Our results provide direct evidence for the hierarchical nature of the representation of visual objects along the visual stream and open up future applications of frequency-tagging methods in fMRI.

Inferring learning rules from distributions of firing rates in cortical neurons

Information about external stimuli is thought to be stored in cortical circuits through experience-dependent modifications of synaptic connectivity. These modifications of network connectivity should lead to changes in neuronal activity as a particular stimulus is repeatedly encountered. Here we ask what plasticity rules are consistent with the differences in the statistics of the visual response to novel and familiar stimuli in inferior temporal cortex, an area underlying visual object recognition. We introduce a method that allows one to infer the dependence of the presumptive learning rule on postsynaptic firing rate, and we show that the inferred learning rule exhibits depression for low postsynaptic rates and potentiation for high rates. The threshold separating depression from potentiation is strongly correlated with both mean and s.d. of the firing rate distribution. Finally, we show that network models implementing a rule extracted from data show stable learning dynamics and lead to sparser representations of stimuli.

Comparison of Object Recognition Behavior in Human and Monkey

Although the rhesus monkey is used widely as an animal model of human visual processing, it is not known whether invariant visual object recognition behavior is quantitatively comparable across monkeys and humans. To address this question, we systematically compared the core object recognition behavior of two monkeys with that of human subjects. To test true object recognition behavior (rather than image matching), we generated several thousand naturalistic synthetic images of 24 basic-level objects with high variation in viewing parameters and image background. Monkeys were trained to perform binary object recognition tasks on a match-to-sample paradigm. Data from 605 human subjects performing the same tasks on Mechanical Turk were aggregated to characterize "pooled human" object recognition behavior, as well as 33 separate Mechanical Turk subjects to characterize individual human subject behavior. Our results show that monkeys learn each new object in a few days, after which they not only match mean human performance but show a pattern of object confusion that is highly correlated with pooled human confusion patterns and is statistically indistinguishable from individual human subjects. Importantly, this shared human and monkey pattern of 3D object confusion is not shared with low-level visual representations (pixels, V1+; models of the retina and primary visual cortex) but is shared with a state-of-the-art computer vision feature representation. Together, these results are consistent with the hypothesis that rhesus monkeys and humans share a common neural shape representation that directly supports object perception.

SIGNIFICANCE STATEMENT

To date, several mammalian species have shown promise as animal models for studying the neural mechanisms underlying high-level visual processing in humans. In light of this diversity, making tight comparisons between nonhuman and human primates is particularly critical in determining the best use of nonhuman primates to further the goal of the field of translating knowledge gained from animal models to humans. To the best of our knowledge, this study is the first systematic attempt at comparing a high-level visual behavior of humans and macaque monkeys.

Deep Neural Networks: A New Framework for Modeling Biological Vision and Brain Information Processing

Recent advances in neural network modeling have enabled major strides in computer vision and other artificial intelligence applications. Human-level visual recognition abilities are coming within reach of artificial systems. Artificial neural networks are inspired by the brain, and their computations could be implemented in biological neurons. Convolutional feedforward networks, which now dominate computer vision, take further inspiration from the architecture of the primate visual hierarchy. However, the current models are designed with engineering goals, not to model brain computations. Nevertheless, initial studies comparing internal representations between these models and primate brains find surprisingly similar representational spaces. With human-level performance no longer out of reach, we are entering an exciting new era, in which we will be able to build biologically faithful feedforward and recurrent computational models of how biological brains perform high-level feats of intelligence, including vision.